[Organizational AI Utilization #149] Turning website information into AI knowledge. The potential of 'Custom AI' created through NotebookLM and Gem integration

Hello! This is Terada.

Currently, I serve as the Representative Director of the AI Digital Community (ADC), a community for AI practitioners such as AI promoters at digital-related companies as well as the Representative Director of FURIKAKE Partners Inc., which supports the 'xAI transformation' of client businesses, and AI Portalize Inc., which provides products that support organizational AI utilization. I support organizational AI utilization from various perspectives! Also, until recently, I served as the Representative Director of D-Marketing Academy Inc., a corporate generative AI & digital marketing talent development training service, and promoted organizational AI utilization at the CARTA HOLDINGS AI Promotion Office. In short, I have been deeply immersed in supporting practical AI utilization for the past two or three years.

Recently, the topic of 'NotebookLM' has been increasing in various places. New features such as slide generation, data table generation, and audio generation are being added one after another, and I often see ways to utilize them being brought up.

Among them, there was an update recently that holds great potential.
It is now possible to link NotebookLM as a source for 'Gem,' which is a custom AI feature, and I would like to write about the potential of how much of a major change this will bring to practical work.

Until now, there was 'instability' when using Web information as a source

Up until now, it was technically possible to have AI read information from websites. This was done by including a URL in the prompt and having it answer based on that content. However, in reality, there was a major issue with the 'stability' of the reading.

I had the impression that it could often only pick up a very small portion of the information on a site, or that it could not process things well when multiple URLs were provided. It was actually difficult to firmly establish information on websites as reliable knowledge for AI.

For example, even if you used custom AI to create something that analyzes the content of a certain URL, there were many cases until now where it would not behave well if you increased the number of URLs too much, or where you had to write an enormous amount of prompts just to acquire the information in the first place.

Incorporating Web information into sources

With this update, this issue is heading toward a resolution all at once. By specifying NotebookLM as a source, it has become possible to stably provide specific Web information to AI. In a word, it feels like you can now install specialized knowledge from the Web directly into the AI's 'brain'.

For example, it is an image of incorporating the following types of information as sources.

• Highly specialized and complex legal rules such as the Pharmaceutical and Medical Device Act and the Act against Unjustifiable Premiums and Misleading Representations

• Unique know-how sites that are systematized within your company or industry

• The latest official documentation with high update frequency, such as GAS (Google Apps Script)

I think that being able to treat these as reliable sources will be a very powerful weapon in practical work.

The difference between NotebookLM alone and Gem integration

Methods such as 'Why not just put information into NotebookLM and ask questions there?' are being introduced on X, YouTube, and other platforms, but there is an aspect where that alone limits it to certain ways of usage.

Of course, it is convenient enough just to extract information, but the biggest merit of linking it with custom AI (Gem) is that you can dramatically stabilize the 'behavior' of the AI.

※Please also see this article regarding the operation of how to divide NotebookLM notebooks.

If we organize the difference between NotebookLM alone and Gem integration, it is as follows.

• NotebookLM alone: Ask questions to the loaded source each time to extract answers.

• Gem integration: After providing reliable knowledge, fix the 'role' and 'behavior' via prompts.

For example, let's consider a case where you load official references such as GAS, GA4, or information from various media outlets. While standard AI might mix in outdated information, you can fix the latest documents in NotebookLM and further define the custom AI (Gem) to 'explain things in an easy-to-understand way for beginners, including code examples.'

By doing this, you can create a highly accurate, dedicated AI bot that possesses both 'limited, up-to-date, and accurate knowledge' and 'the precise behavior you intended.'

※Please also see this article regarding what to put into the 'brains' of NotebookLM and custom AI (Gem).

Using NotebookLM as a source to make Gems more flexible

I feel that this integration has significantly raised the customizability of AI. In situations where we previously had to rely on search-based AI, we can now leave it to a 'dedicated AI customized to our own preferences.'

Why not start by putting the professional sites you frequently refer to in your work or the latest references into NotebookLM and creating your own dedicated Gems? It should be a big step toward evolving AI from a 'tool for searching' into a 'partner with specialized knowledge.'

Please try it out in your organization's AI utilization.

※However, updates to NotebookLM sources or information updates when the source information itself changes are not done automatically. Please refer to the article below regarding that aspect.

Thank you for reading!
Click here for past AI-related articles!

By the way, I have also published a book summarizing AI promotion in organizations. Please take a look if you are interested.